Skip to content

Bound native CUDA cache recovery by measured forward work #900

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
bradhilton merged 6 commits into from
Sep 17, 2026
Merged

Bound native CUDA cache recovery by measured forward work #900

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
2b84bc9
Bound native cache recovery by measured forward work
bradhilton Sep 15, 2026
ce72470
Integrate current planning checks and preserve primary recovery errors
bradhilton Sep 16, 2026
be74fa3
Ensure checkpoint slots before DP-local recovery retries
bradhilton Sep 16, 2026
75e18b8
Type the scalar recovery-plan test double explicitly
bradhilton Sep 16, 2026
40cb87b
Merge current main into bounded cache recovery and cover dense CP adm…
bradhilton Sep 17, 2026
7480edf
Fix cache recovery test metadata and CI routing
bradhilton Sep 17, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions .github/workflows/prek.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -226,6 +226,8 @@ jobs:
tests/unit/test_trainer_rank_validation.py \
tests/unit/test_trainer_rank_weird_shapes.py \
tests/unit/test_trainer_rank_split.py \
tests/unit/test_trainer_rank_cache_recovery.py::test_dense_cp_exact_demand_fits_after_recovery \
tests/unit/test_trainer_rank_cache_recovery.py::test_dense_cp_exact_demand_refuses_after_recovery \
tests/acceptance/trainer_rank_planner \
tests/integration/megatron/test_sft_packing.py::test_sft_packing_preserves_training_targets \
tests/integration/megatron/model_support/test_dispatcher_graph_retention.py \
Expand All @@ -240,6 +242,8 @@ jobs:
- name: Run unit tests
run: |
uv run --no-sync pytest --nbval --current-env --tb=short tests/unit \
--deselect=tests/unit/test_trainer_rank_cache_recovery.py::test_dense_cp_exact_demand_fits_after_recovery \
--deselect=tests/unit/test_trainer_rank_cache_recovery.py::test_dense_cp_exact_demand_refuses_after_recovery \
--ignore=tests/unit/test_megatron_reference_logprobs.py \
--ignore=tests/unit/test_moe_routing_replay.py \
--ignore=tests/unit/test_moe_routing_real_path.py \
Expand Down
547 changes: 445 additions & 102 deletions src/art/trainer_rank/_impl.py
View file
Open in desktop

Large diffs are not rendered by default.

6 changes: 3 additions & 3 deletions tests/unit/test_trainer_rank_active_memory.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -187,16 +187,16 @@ def test_warm_admission_rechecks_current_residency(monkeypatch):
)
monkeypatch.delenv("ART_TRAINER_RANK_TEST_HOOKS", raising=False)
# Fresh memory accounting observes newly resident state without discarding
# a valid incremental profile; cached free blocks remain reusable.
# a valid incremental profile; cache alone is not physical availability.
for allocated, reserved, fits in [
(10_000, 10_000, True),
(90_000, 90_000, False),
(10_000, 90_000, True),
(10_000, 90_000, False),
]:
state.update(allocated=allocated, reserved=reserved)
check = rank._memory_check(plan)
assert check.estimated_required_bytes == required
assert check.available_bytes == total - allocated - 3000
assert check.available_bytes == total - reserved - 3000
assert check.fits == fits
assert rank._memory_profiles[plan.signature] == profile

Expand Down
Loading
Loading